We changed the agent: did it actually get better? Every self‑improving agent loop asks this hundreds of times, and each answer comes from a verifier. For open‑ended tasks where no ground‑truth evaluator exists, the loop is forced to rely on a hand‑written rubric or a bare LLM judge that grades the output of a model like itself, opening the door to reward hacking and shared blind spots. We make the verifier the evolving object: an inspectable expression built from a small set of mostly deterministic drawback detectors, synthesized from clustered failures, gated at birth, and selected for agreement with a ten‑item anchored reference set plus consensus over unlabeled outputs—never for the agent’s score.
On the MBPP+ benchmark the verifier gains $+0.21$ held‑out agreement over the hand‑authored seed composition, on every seed, and outperforms the bare LLM judge it contains. One crucial finding is that removing the anchor guards collapses the verifier into a vacuous always‑pass grader, yet that collapsed verifier still trains skills just as well. Downstream task scores cannot certify a self‑evolved verifier; scores only answer sufficiency, and in that sense an evolved verifier can substitute: using a “Double Ratchet” that pairs the verifier with a lifecycle‑managed skill loop retains $88\%\sim110\%$ of the lift that ground truth or a rubric provides, across code generation, enterprise text‑to‑SQL, and reference‑free report generation.
When evolved skills gamed the report rubric, an outer judge caught the abuse and an added detector repaired it; the judge itself was wrong until given the task contract.
Review: The paper demonstrates that treating the verifier as a co‑evolving, inspectable component can markedly improve the reliability and generality of self‑improving agents, especially in open‑ended settings lacking clear evaluation criteria, offering a promising new direction for robust AI alignment.